Positive-Sum AI: earning permission for the agentic future

Artificial intelligence may create enormous value. That does not mean people will permit it to reshape their lives. A technology can be productive, profitable, scientifically impressive and strategically important — and still lose its legitimacy.

It can create more total wealth while making workers less secure, communities more burdened, markets more concentrated, citizens less autonomous, countries more dependent and customers easier to manipulate. It can create value in aggregate while distributing gains upward, costs outward, risks downward, and responsibility nowhere.

Which is why the defining question of the AI era is not how much economic value artificial intelligence will create.

Who receives that value, who bears the transition costs, who retains agency — and what evidence shows the result is better for society rather than merely more profitable for its owners?

Positive-sum AI is described far too casually. A company launches an AI product. Productivity increases. Management declares that everyone will eventually benefit. The conclusion does not follow. Productivity creates the possibility of shared prosperity; it does not determine wages, employment, working hours, consumer prices, market concentration, tax contributions, environmental burdens or access to opportunity. Those are shaped by ownership, competition, bargaining power, institutional design, corporate choices, public policy and social norms.

Positive-sum AI is therefore not a property of the model. It is a property of the complete system built around it — and a model cannot make a transformation positive sum by itself. An organization can. A society can. Or they can fail to.

Satya Nadella’s argument, in the conversation this series has been drawing on, is unusually direct. The technology industry cannot continue telling people that white-collar work may disappear, that they should nevertheless welcome the technology, and that unspecified future jobs will somehow emerge. Communities hosting data centres must see genuine benefits — schools, tax base, energy, water, infrastructure, local prosperity. Workers need credible answers about which new jobs will exist, what they will pay, and how to qualify for them. Companies and countries must not become mere suppliers of data to systems whose economic value is captured elsewhere.

Imagine a company that deploys AI across its customer operations. Within two years, customer-response time falls by 70%, operating cost falls by 35%, revenue per employee rises, and profit reaches a record level. The investor presentation describes a successful AI transformation.

Now read the same transformation from five ledgers.

The experienced employee is producing more value and capturing little of it. The junior role has disappeared, and the company promises future jobs in agent supervision, workflow design and evaluation — jobs that require experience the junior worker has not yet acquired. The theoretical new opportunity exists. The pathway does not.

Was the transformation positive sum? The answer cannot be determined from profit or productivity alone, because the company also changed distribution, opportunity, control, public infrastructure and bargaining power.

what organizations calculate

    benefits to the firm − costs to the firm = value created


what the complete calculation looks like

    benefits to all affected stakeholders
  − financial, human, social, political and environmental costs
  ± adjustment for distribution, agency, durability and transition

The second calculation is difficult. That does not make it optional.

A positive-sum transformation creates more total value than it destroys. It does not automatically mean every person benefits. Technological transitions can produce large aggregate gains and severe concentrated losses; both statements can be true at once.

A strict Pareto improvement — someone better off, nobody worse off — is a fantasy at this scale. AI will alter occupations, reduce demand for some skills, weaken some business models, shift investment, change communities and create new forms of power. There will be losses. Pretending otherwise weakens trust rather than building it.

Economists sometimes call a change beneficial when winners could compensate losers and remain better off. But could compensate is not did compensate. The phrase “society will benefit overall” often conceals the absence of any mechanism through which displaced workers, affected communities, excluded countries or disrupted suppliers actually share in the gain.

  durable additional value
+ distributed opportunity
+ human capability
+ public and environmental benefit
− displacement cost
− externalized harm
− concentration risk
− loss of agency
─────────────────────────────
= positive-sum value

A transformation should not be called positive sum merely because one stakeholder gains enough to outweigh many smaller losses mathematically. A credible standard also asks whether basic rights are protected, whether some stakeholders carry unacceptable burdens, whether transition pathways are real, whether affected people can participate in decisions, whether harms are reversible, and whether future generations are being charged for current returns.

A company may hold permits, contracts, regulatory approval and technical capability — and still face worker resistance, customer distrust, community opposition and political backlash. That is the difference between legal authority and social permission.

Public concern here is not ignorance. In a 2025 Pew survey, 52% of US workers said they were worried about future workplace AI use, while only 6% expected it to create more job opportunities for them personally. In another survey, the US public was much more likely than AI experts to expect personal harm, job loss and negative societal effects — though both groups expressed strong interest in greater control over how AI is used. These views may change. They should not be dismissed as resistance to progress; they reflect rational questions about control, employment, distribution and trust.

Permission grows when people can point to outcomes: a safer job, higher wages, reduced working hours, improved healthcare, accessible education, stronger local infrastructure, new business creation, lower consumer prices, meaningful environmental improvement. And the proof must be tangible, attributable, understandable and sustained.

Productivity is the economic promise of AI and the primary mechanism through which it may raise living standards. Suppose a team produces the same output in half the time. That surplus can become more products, lower prices, higher quality, greater profit, higher wages, shorter hours, new investment — or fewer employees.

The technology does not choose among them. The institution does.

When productivity rises and bargaining institutions remain unchanged, gains may accrue disproportionately to capital owners, scarce technical talent and dominant platforms. An IMF model published in 2025 highlights that AI’s effect on inequality may operate through several competing channels: wage inequality could narrow if high-income work is displaced, but complementarity with highly skilled workers and their stronger ownership of capital may increase both income and wealth advantages. Positive-sum distribution cannot be assumed.

Mechanisms for sharing the dividend with workers include gain-sharing bonuses, wage progression, employee ownership, reduced hours, learning accounts, internal mobility guarantees, transition income and new-role pathways. No single mechanism fits every company. The absence of any mechanism is itself a choice.

One of the most underused dividends is time. If AI reduces low-value work, the organization can convert the gain into deeper customer work, learning, innovation, reduced overload or shorter workweeks. Instead, companies may simply raise output expectations. The employee becomes more productive and more exhausted. This is not the only possible future.

Not all AI that helps a company helps its workers. A system may increase output by monitoring employees more intensely, simplifying jobs, transferring judgment to management, reducing staffing and accelerating work. That is productive AI. It is not necessarily pro-worker AI.

Research by Daron Acemoglu, David Autor and Simon Johnson distinguishes several ways technology can interact with labour, and argues that new-task creation is the most unambiguously pro-worker because it creates fresh demand for human expertise rather than making existing labour cheaper or more replaceable.

Without those nine answers, “new jobs will emerge” is not a plan. It is an abstraction.

The ILO’s refined 2025 global exposure index estimates that roughly one in four jobs has some exposure to generative AI, while emphasizing that transformation of task bundles is more likely than wholesale automation of most occupations — and that outcomes depend heavily on implementation, consultation and institutional context. Transformation can still be painful: reduced hiring, role consolidation, wage pressure, fewer entry points, greater performance expectations. Positive-sum design must address these before they appear in employment statistics.

An employee may be told: use this tool, accept this score, meet the new productivity target, trust the system, retrain yourself. That is deployment without agency. Real agency includes information, participation, influence, challenge, appeal, skill development and meaningful choice.

Consultation is not a courtesy; it is productive infrastructure. ILO case studies across multiple regions show that social dialogue can support a “high road” approach in which AI complements skills, empowers workers and embeds new work within labour and social protections rather than merely replacing or controlling employees. It has to begin before procurement, before workflow design, before metrics are selected, and before staffing decisions.

There is also a right not to be invisibly managed. AI systems should not make consequential workforce judgments through opaque inferences about loyalty, emotion, personality, intent or future performance. Human beings are not data points inside an optimization system.

A job is more than income

Work provides security, autonomy, identity, mastery, social connection, purpose and progression — and AI can improve or weaken each. It may remove administrative burden, improve safety, support disabled workers, reduce stress, increase flexibility and expand capability. A 2026 study of deaf and hard-of-hearing delivery workers found that an AI communication system improved speed, customer ratings, labour supply and retention in the studied setting, illustrating how AI can remove barriers rather than merely automate labour. It may equally intensify pace, monitor continuously, fragment tasks, reduce discretion, eliminate interpersonal contact and create permanent availability.

This is the augmentation trap. An employee receives a powerful assistant. Management then increases expected output by more than the assistant’s actual contribution. The employee experiences greater work intensity, more responsibility and less recovery time. A productivity improvement that degrades most of these dimensions should not be described as unambiguously positive sum.

Economic theory may predict new jobs and higher productivity over time. A displaced worker lives in a specific city, with specific skills, financial commitments, a family and limited time.

The long run does not pay next month’s rent.

Transition costs are real costs: income loss, retraining time, geographic relocation, identity loss, health effects, family disruption, career scarring.

And training cannot solve every displacement. Some people lack time, face health constraints, are near retirement, live far from new jobs, or carry caregiving responsibilities. A positive-sum transition includes social protection, not only education.

AI can increase conversion, personalization, retention and usage. These may benefit the company without benefiting the customer.

The more a system knows about a person, the better it may serve them, predict them and influence them. Positive-sum design preserves the distinction between assistance and exploitation. Customers should understand, where appropriate, that they are interacting with AI, what the system can decide, how data is used, how to reach a human and how to challenge an outcome.

Stronger protections may be needed for children, elderly people, financially distressed customers, people with disabilities, and people seeking health or legal help.

Children are not simply smaller users

They have developing judgment, weaker capacity to detect persuasion, long-term developmental needs, and special privacy and safety interests. Positive-sum AI for children should strengthen curiosity, critical thinking, creativity, physical-world exploration, relationships and cognitive coverage. It should not merely make homework faster. Educational value depends on whether children understand more, develop capability, remain curious and retain independent thought.

Current AI development consumes energy, water, public infrastructure and social attention. Future generations should inherit knowledge, productive capacity, a stable climate and functioning institutions — not only larger models.

A platform may succeed by enabling developers, suppliers, partners and customers to create value. Or it may subsidize participation, establish dependency, capture the profitable layer, and compete against its own partners.

Nadella’s argument is that long-term platform stability comes from ensuring the value created on top of the platform substantially exceeds the value captured by the platform itself. Partners trust the system when it is not a disguised zero-sum strategy designed eventually to eat their lunch. The ratio need not be exact; the discipline matters.

The same logic applies to data. Companies may provide proprietary data, employee feedback, customer interactions and workflow traces that improve an external model ecosystem. If they retain no meaningful learning, bargaining power, economic return or portability, the arrangement is extractive — the corporate case that Token Capital and AI Sovereignty both made. Positive-sum partnership requires clear rights, transparent economics, portability, shared innovation, non-retaliatory exit and appropriate value sharing.

And large companies should ask whether AI-driven procurement expands supplier access, reduces administrative burden and improves payment — or merely increases price pressure, automates exclusion and transfers compliance cost onto small suppliers.

AI is not weightless. It depends on land, electricity, grids, water, construction, chips, data centres and public permitting. The costs are frequently local. The benefits may be global.

The IEA’s 2026 analysis states that community opposition, electricity affordability, environmental impacts, grid connections, chip supply and planning capacity are becoming material constraints on AI-infrastructure expansion — and projects data-centre electricity consumption rising from about 485 TWh in 2025 to roughly 950 TWh by 2030 in its central case.

If public systems must absorb the cost of transmission, roads, water and emergency capacity, those costs belong in the project’s economic assessment. A community needs data, enforceable commitments, participation, grievance channels and long-term accountability.

Energy and climate

AI has a dual environmental role: it consumes energy, and it may improve grids, industry, buildings, transport and scientific discovery. But efficiency per task can improve rapidly while total consumption still rises, because usage expands, models become more capable, reasoning grows longer, and video and agentic workloads grow. The IEA reports that energy use per simple AI task has fallen sharply while more energy-intensive reasoning, agentic and media-generation uses expand.

Abundance can coexist with concentration. AI may make intelligence cheaper and more widely available while the infrastructure enabling it remains concentrated — in advanced chips, fabrication, cloud, frontier models, distribution platforms, proprietary data and capital requirements. OECD analysis identifies capital intensity, scale economies, supply-chain bottlenecks and vertical relationships as potential competition concerns across AI infrastructure.

Competition matters for positive sum because it supports lower prices, innovation, choice, bargaining power and wider value distribution. And access is not the same as leverage: a company may have access to an AI service while lacking portability, negotiating leverage, control of learning, or the ability to compete with its provider. A platform that simultaneously provides infrastructure, controls distribution, observes customer behaviour and competes with its customers holds a structural advantage.

AI can genuinely democratize capability — a small team gaining access to software development, design, analysis, translation, customer support, legal preparation and scientific tools; a founder operating globally using agents. That also brings high market volatility, weak institutional safeguards, extreme concentration of founder authority and fewer traditional jobs.

Meanwhile OECD data for 2025 show AI use remaining much higher among large firms than small firms even as overall adoption increases, raising the risk that AI widens productivity differences between already capable organizations and smaller businesses lacking data, skills, capital and integration support. Positive-sum small-business policy may include shared compute, open tools, technical assistance, interoperable standards, public evaluation resources, SME financing and sector data spaces — with the objective of increasing productive autonomy, market access and capability, not merely converting every small company into another firm’s platform tenant.

Countries differ in electricity, connectivity, compute, capital, data, language resources, education and institutional capacity. IMF modelling suggests AI may increase cross-country inequality because advanced economies combine higher exposure with greater preparedness and access to essential technologies; projected growth effects may be more than twice as large in advanced economies as in low-income countries under the paper’s assumptions.

A 2026 ILO–World Bank analysis warns that developing economies may experience labour-market disruption before receiving comparable productivity benefits, because of gaps in digital infrastructure and differences in task composition. That is a deeply negative-sum possibility for the countries concerned: work is displaced globally, value is captured elsewhere, and domestic productivity does not rise sufficiently to compensate.

Language exclusion becomes economic exclusion. And a country or community may contribute language, cultural knowledge, user behaviour and public data while receiving little ownership, infrastructure, capability or economic return — the pattern usually named data colonialism.

AI can expand state capacity across benefit administration, tax service, permitting, healthcare, education, fraud detection, legal access and emergency response. The public-service test is whether a system improves access, speed, consistency, fairness and human service quality without weakening due process, explanation, appeal, equal treatment and democratic accountability.

An automated benefit system may reduce administrative cost while incorrectly excluding vulnerable people. The saving appears in one budget. The harm appears in another life.

Public AI cannot optimize only for throughput; it must include rights, equity, accessibility, public trust and error remedy. Governments can also support positive-sum AI directly through digital identity, open data, shared compute, public research, interoperability and procurement standards.

Healthcare, science, and human capability

The strongest social case for AI may be capability expansion: understanding disease, discovering materials, designing treatments, improving diagnosis, accelerating science. Nadella notes that social permission might have developed differently if the earliest widely visible AI breakthroughs had centred on scientific discoveries with obvious societal benefit rather than primarily on job-replacement narratives.

But access matters: a medical capability that exists while remaining unaffordable, geographically inaccessible or restricted to premium systems does not fulfil its positive-sum potential. AI should support diagnosis, administration and monitoring in order to create more room for attention, empathy, relationship and informed consent — not less. And since publicly funded research, open datasets and shared scientific knowledge contribute to AI capability, the resulting benefits should not be enclosed so completely that the public pays twice: once for the knowledge, again for inaccessible outcomes.

Education

In a world of abundant machine-generated answers, education must place greater value on questions, judgment, verification, synthesis, curiosity, collaboration and cognitive coverage. AI can make high-quality educational support available to far more people. It can also let students outsource writing, problem-solving, reflection and practice — achieving higher short-term performance with weaker learning.

Positive-sum educational AI therefore measures retained knowledge, transfer, curiosity, independent performance, confidence and access — not merely task completion. It should strengthen teachers’ ability to understand students, prepare instruction and give feedback, rather than converting education into automated content delivery with minimal human relationship. And if AI performs many of the outputs used to assess knowledge, institutions must redesign examinations, portfolios, live demonstrations, apprenticeships and collaborative assessment.

AI can improve democratic participation through translation, policy understanding, public consultation, access to law and government transparency. It can also industrialize persuasion: personalized propaganda, synthetic identities, deepfakes, automated harassment, large-scale misinformation.

A highly resourced actor can use AI to test messages, personalize influence and operate continuously. The citizen has limited attention. This is not a balanced interaction, and citizens should not become objects optimized by invisible persuasion systems. A positive-sum information ecosystem requires provenance, authentication, pluralism, media literacy, transparency and institutional resilience — because the entire positive-sum case for AI collapses if people can no longer trust what they see, who they are speaking with, or whether institutions are authentic.

The distribution of AI gains is strongly influenced by who owns the models, compute, platforms, data, companies and token capital. Broadening that ownership is possible through employee equity, pension-fund participation, public investment vehicles, cooperative platforms, shared intellectual-property structures and community ownership of infrastructure.

Employees contribute corrections, examples, tacit knowledge, decision patterns and customer understanding — the raw material of token capital. Companies should decide how that contribution is recognized, compensated, governed and protected. Individuals and communities contribute data through mechanisms that could include consent, data trusts, collective licensing, public-interest conditions and benefit sharing.

Productivity also changes the tax base. If production shifts from labour toward software, compute, intellectual property and capital, tax systems relying heavily on labour income come under pressure — and tax neutrality is not always neutral, since systems may implicitly favour automation investment and capital equipment over employment, training and worker development. Possible directions include reforming capital and corporate taxation, reducing penalties on employment, providing training credits, funding transition systems, taxing rents rather than productive adoption, and strengthening international coordination.

AI relies on public inputs — education, research, infrastructure, legal systems, energy networks — so a portion of exceptional gains can legitimately support transition, public services, infrastructure and broad capability. A simplistic “robot tax” is the wrong instrument: hard to define, and liable to discourage beneficial productivity. The deeper goal is to tax economic rents appropriately, preserve public revenue, and avoid subsidizing socially harmful substitution.

Intent is not evidence.

Do not collapse these into one opaque score too quickly. A high total can conceal a severe rights violation, an unacceptable local harm or a critical safety risk. Some thresholds must be non-compensable: unlawful discrimination, unsafe product release, coercive worker surveillance, unauthorized data use, denial of meaningful appeal. A benefit elsewhere does not cancel these.

Before deploying a material AI transformation, leadership can publish a covenant — ten questions answered in public, in advance.

A major FMCG company introduces an agentic field-sales and retail-execution system serving 80,000 independent retailers through 1,200 field-sales representatives across several national markets. The system can prioritize stores, recommend orders, identify assortment gaps, generate visit plans, monitor promotions, automate reporting and suggest sales conversations.

The narrow automation case assumes an 18% productivity improvement, a 15% reduction in field headcount and improved promotional compliance. The project appears attractive. Then read it from the other ledgers.

The redesign begins by changing the objective. From:

before   maximize revenue per field employee

after    improve retailer growth, consumer availability, field
         productivity, employee capability and route sustainability
         while preserving local commercial judgment

The company commits part of verified productivity gains to capability bonuses, training, reduced administrative load, internal mobility and entrepreneurship support. It does not promise zero workforce change. It defines a transition path.

Retailers can reject recommendations, explain local constraints, request human support and understand why an assortment is proposed. Participating retailers receive simple demand insights, inventory recommendations, digital-ordering tools and training — so the system creates capability for the retailer rather than merely extracting more sell-in. The company limits location tracking outside work, behavioural inference and automated disciplinary use; performance data supports coaching before punishment. Route optimization reduces driving distance, fuel and failed visits, and the company measures the actual reduction rather than modelling theoretical savings.

The company may still operate with fewer employees over time. The transformation becomes more positive sum because workers share gains, transition is supported, remaining jobs improve, retailers receive capability, human judgment remains, environmental outcomes are measured, and affected stakeholders can challenge the system.

Positive sum does not mean no change. It means change designed to create more value than harm — and to distribute the value credibly.

1. 01What additional value does our AI strategy create?
2. 02Which stakeholders receive that value?
3. 03Which stakeholders bear the greatest costs?
4. 04How much of the productivity dividend goes to workers?
5. 05Which current jobs may disappear or diminish?
6. 06What specific new roles will exist?
7. 07What will those roles pay?
8. 08Who can realistically transition into them?
9. 09How are worker representatives involved?
10. 10Are remaining jobs becoming better, or merely more intense?
11. 11Do customers receive lower prices, better quality, or greater access?
12. 12Can customers reach a human and appeal decisions?
13. 13What infrastructure costs do communities bear?
14. 14How are local electricity, water and public-service effects managed?
15. 15Are we strengthening or weakening competition?
16. 16Do partners create more value than our platform captures?
17. 17Are we accumulating proprietary capability, or feeding another firm's intelligence?
18. 18How are global and language inequalities affected?
19. 19Which benefits are measured outcomes, and which remain promises?
20. 20What evidence would cause us to conclude that the transformation is not positive sum?

AI capability:                    Business objective:
Additional value created:         Baseline:
Company benefit:                  Worker benefit:
Customer benefit:                 Partner benefit:
Community benefit:                Public benefit:
Environmental benefit:            Affected roles:
Transition cost:                  Externalities:
Distribution mechanism:           Agency safeguards:
Competition impact:               Key evidence:
Decision:

Workflow:                         Productivity gain:
Verified economic value:
Share to investment:              Share to workers:
Share to customers:               Share to communities/public:
Share to shareholders:
Mechanism:                        Review date:

Role affected:                    Number of people:
Timing:                           Tasks changing:
Future roles:                     Required skills:
Training pathway:                 Income protection:
Internal mobility:                External pathway:
Employee consultation:            Owner:

Facility or deployment:           Community:
Local jobs:                       Tax contribution:
Grid requirement:                 Ratepayer impact:
Water use:                        Environmental impact:
Public infrastructure:            Local procurement:
Community commitments:            Grievance process:
Independent verification:

AI interaction:                   Customer benefit:
Data used:                        Decision made:
Disclosure:                       Human alternative:
Appeal:                           Vulnerable groups:
Manipulation risk:                Error remedy:
Owner:

Platform or capability:           Partners:
Value created by ecosystem:       Value captured by platform:
Data contributed:                 Learning retained by partners:
Portability:                      Competing-provider risk:
Contract fairness:                Improvement action:

Period:                           Value created:
Value distributed:                Workers displaced:
Workers transitioned:             Job-quality change:
Customer-surplus change:          Community impact:
Environmental impact:             Competition impact:
Agency incidents:                 Targets missed:
Corrective actions:

What is positive-sum AI?

AI that creates additional durable value, distributes meaningful benefits, protects human agency and rights, supports those bearing transition costs, and preserves the conditions required for future prosperity.

Is all productivity-enhancing AI positive sum?

No. Productivity may rise while wages stagnate, jobs worsen, market power increases, communities bear costs and customers lose agency.

Does positive sum mean no one loses?

No. Major transformations often create concentrated losses. Positive-sum design requires honest recognition, transition support, compensation where appropriate, and protection of rights.

Is social permission the same as regulatory approval?

No. An organization can be legally permitted while lacking public trust or stakeholder acceptance.

Who should receive the productivity dividend?

The distribution is a design and policy choice. Potential beneficiaries include workers, customers, shareholders, partners, communities and government.

Is job transformation always better than job loss?

Not necessarily. A transformed job may involve lower autonomy, greater monitoring, work intensification or reduced pay. Job quality must be measured.

Will AI create new jobs?

It likely will create some new tasks and roles. The relevant questions are how many, where, at what wage, accessible to whom, and on what timeline.

Is retraining enough?

No. Retraining must connect to actual employment demand, and may need to be combined with income support, mobility and social protection.

Should companies guarantee no layoffs from AI?

Some may choose time-bound guarantees; a universal guarantee may not be practical. Every company should provide transparency, fair process, transition support and credible value-sharing mechanisms.

Can reducing working hours be an AI dividend?

Yes. Productivity can be converted into more output, fewer employees, shorter hours or better work. The choice is institutional.

What is extractive customer AI?

AI that primarily increases the company’s ability to manipulate, discriminate, lock in or capture value from customers without proportionate benefit.

Is AI environmentally positive or negative?

It can be either. AI consumes physical resources and may also enable environmental improvements. Both must be measured, including rebound effects.

Does every country need its own frontier model?

No. Countries need sufficient capability to adapt, govern, choose, create local value and preserve strategic options.

Can positive-sum AI be measured with one number?

Not credibly in most cases. Several dimensions should remain visible, and some rights and safety thresholds should not be traded away against benefits elsewhere.

Who owns positive-sum outcomes in a company?

The CEO and board own the overall institutional commitment. Business, HR, technology, finance, sustainability, legal, risk and workforce representatives share execution responsibility.

What is the first practical step?

Take one major AI business case and add four missing columns: who benefits, who bears cost, how gains are distributed, and what evidence will prove the outcome.

Artificial intelligence may be the most powerful general-purpose technology of our era. It may help humanity discover medicines, improve education, make work safer, increase productivity, create new companies, reduce waste, expand access to expertise and understand complex systems. It may also displace work, intensify surveillance, concentrate ownership, weaken local economies, strain physical infrastructure, reduce human agency and widen global inequality.

Both futures are technologically possible. Neither is guaranteed. That is the most important fact in this essay.

The AI future will not be determined by model capability alone. It will be determined by corporate strategy, labour institutions, ownership, competition, taxation, education, infrastructure, democratic governance and human choices.

So positive sum should never become a ceremonial adjective attached to every AI initiative. It should be a demanding standard. A company should not say this is positive sum because the technology creates growth. It should be able to show which value was created, who received it, which costs were created, who carried them, and what happened to worker capability, customers, communities, competition and human agency.

Positive sum is not a forecast. It is an operating model.

It begins by rejecting two comforting stories.

The company of the agentic age must therefore accept a new responsibility. It is no longer enough to build systems that are accurate, secure, compliant and profitable. They must also be designed to create a legitimate place for workers, customers, communities, partners and future generations. This does not require companies to solve every social problem. It requires them to stop pretending their choices are socially neutral.

When a company decides which tasks to automate, which workers to train, which data centre to build, which customer behaviour to optimize, which model provider to depend on, and how to distribute productivity gains — it is designing part of the political economy of the AI age.

The technology industry cannot earn permission by asking society to trust its intentions. Companies cannot earn permission by publishing principles while privatizing the gains and socializing the costs. Leaders cannot earn permission by promising jobs that have not been specified, prosperity that has not been distributed, or environmental benefits that have not been measured.

Permission will come from proof.

• That workers became more capable
• That customers received real value
• That displaced people received real pathways
• That communities became stronger
• That small firms gained opportunity
• That science advanced
• That public services improved
• That infrastructure remained sustainable
• That human agency was preserved

The age of AI will not retain legitimacy because progress is inevitable. It will retain legitimacy only if progress becomes recognizable in people’s lives. That is the final challenge of the agentic firm: not merely to build intelligence, not merely to govern agents, not merely to accumulate token capital — but to build an institution whose growing intelligence makes the world around it more capable too.

That is how the agentic firm earns permission to exist. That is how AI becomes positive sum. And that is how we preserve the right to build the future.

The source conversation. Grounded in the supplied Satya Nadella–Reid Hoffman discussion, particularly its call for tangible societal proof points, real employment pathways, community benefits from AI infrastructure, company and national sovereignty, positive-sum ecosystems, and AI that improves the human condition broadly rather than narrowly.

Public permission and trust. Pew Research findings from 2025 show meaningful gaps between public and expert optimism, substantial worker concern about workplace AI, and broad demand for greater control over how AI is used.

Work transformation.The ILO’s 2025 refined global exposure index concludes that task and job transformation is more likely than wholesale occupational replacement, while stressing that effects vary across countries, sectors, occupations and implementation choices.

Pro-worker AI. Acemoglu, Autor and Johnson distinguish labour augmentation, capital augmentation, automation, expertise levelling and new-task creation, arguing that current market incentives may underproduce AI designed to increase the value of human capability and expertise.

Social dialogue. ILO case studies show how worker representatives and collective institutions can shape AI adoption toward skill complementarity, empowerment, job quality and labour protections.

Productivity and inclusion. Field research provides evidence that generative AI can increase productivity and spread expertise in some work settings, while newer accessibility research shows how appropriately designed AI can improve outcomes for workers with disabilities.

Inequality. IMF research highlights competing wage, wealth, complementarity and capital-ownership channels through which AI adoption could affect inequality within and across countries.

Human development.The UNDP’s 2025 Human Development Report frames AI’s impact as a matter of choices and human possibilities rather than a deterministic consequence of technology.

Global AI divide.World Bank and ILO–World Bank research identifies connectivity, compute, data, skills, infrastructure and task composition as important determinants of whether countries receive AI’s benefits or experience disruption without comparable productivity gains.

Competition. OECD analysis examines how capital intensity, supply-chain bottlenecks, advanced computing, vertical integration and concentrated infrastructure may shape competition and value distribution across the AI economy.

Energy, communities and environmental impact.The IEA’s 2025 and 2026 research documents rapidly growing data-centre electricity demand, increasingly visible community and affordability concerns, infrastructure constraints, efficiency improvements, and the potential for AI to support energy-system improvements while also creating rebound effects.